v0.10.0

Added

• Jina Reader ships as a built-in skill. Mention $jina-reader with a
  public webpage or PDF URL to read it through Jina Reader using curl. It
  covers Markdown, text, HTML, JSON and screenshot links, works without a key
  for basic access, and supports optional JINA_API_KEY authentication. Like
  skill-creator, it is installed on first launch, preserves local edits and
  can be disabled through /skills; custom skill-directory overrides remain
  untouched. Loading the skill itself makes no network request
  (docs/skills.md).

• The agent knows which user it runs as. The environment block in its
  system prompt gains a User line — User root, User linuztx — so it
  knows whether it is root before it reaches for sudo. Root is decided by
  the user id, never by a name: a root account under another name reads
  toor (root), a leftover $USER=root cannot make an ordinary user look
  like root, and an account with no name reads uid 1000. Subagents carry
  the same line (docs/environment.md).

Changed

• Shorter, clearer skill-tool guidance. The description tells the model
  to load a matching or mentioned skill once and reuse its instructions while
  they remain in context. It removes conflicting advice that required another
  tool call whenever mentioning a skill.

Fixed

• Repeated skill calls no longer duplicate instructions in context. The
  executor returns a short "already loaded" reminder without rereading
  SKILL.md or appending its body, even when the model repeats a call within
  the same batch. Detection follows each agent's actual context across turns,
  resumed sessions, and subagent continuations; each new subagent can load
  its own copy. Compaction, clearing, or backtracking permits a fresh load
  when the original instructions are gone. Failed loads remain retryable,
  disabled skills remain unavailable, and name aliases, hook-rewritten calls,
  and secret-masked directory headers are handled (docs/skills.md).
• Each account gets its own temp directory where /proc is missing.
  The session's scratchpad and background output live under
  alter-zero-{uid}, and the uid was read from /proc/self, which answered
  0 wherever /proc is absent: on a Linux sandbox without it every account
  shared /tmp/alter-zero-0, where a second account could not create its
  session. The uid now comes from the kernel's effective uid on every
  platform (docs/background.md).

Full changelog: v0.9.0...v0.10.0